private void SetBackwardPassArgs()
{
// Arg 1: layerNodeCount (uint)
ClHelpers.SetKernelArg(
_backwardKernel,
1,
(uint)NodeCount
);
// Arg 2: nextLayerNodeCount (uint)
ClHelpers.SetKernelArg(
_backwardKernel,
2,
(uint)0
);
// Arg 3: layerSums (float*)
_layerSums.SetKernelArg(_backwardKernel, 3);
// Arg 4: layerOutputs (float*)
_layerOutputs.SetKernelArg(_backwardKernel, 4);
// Arg 5: layerOutputs (float*)
_layerDeltas.SetKernelArg(_backwardKernel, 5);
// Arg 6: layerWeights (float*)
_weights.SetKernelArg(_backwardKernel, 6);
// Arg 7: layerBiases (float*)
_biases.SetKernelArg(_backwardKernel, 7);
// Arg 8: previousLayerOutputs (float*)
GpuMatrix.SetNullKernelArg(_backwardKernel, 8);
// Arg 9: nextLayerDeltas (float*)
GpuMatrix.SetNullKernelArg(_backwardKernel, 9);
// // Arg 10: nextLayerWeights (float*)
GpuMatrix.SetNullKernelArg(_backwardKernel, 10);
// Arg 11: targets (float*)
GpuMatrix.SetNullKernelArg(_backwardKernel, 11);
}
private void SetForwardPassArgs()
{
// Arg 0: previousLayerNodeCount (uint)
ClHelpers.SetKernelArg(
_forwardKernel,
0,
(uint)_previousLayerNodeCount
);
// Arg 1: layerNodeCount (uint)
ClHelpers.SetKernelArg(
_forwardKernel,
1,
(uint)NodeCount
);
// Arg 3: layerWeights (float*)
_weights.SetKernelArg(_forwardKernel, 3);
// Arg 4: layerBiases (float*)
_biases.SetKernelArg(_forwardKernel, 4);
// Arg 5: layerSums (float*)
_layerSums.SetKernelArg(_forwardKernel, 5);
// Arg 6: layerOutputs (float*)
_layerOutputs.SetKernelArg(_forwardKernel, 6);
}
public void SetTargets(ExecutionSequence executionSequence, double[] targets)
{
if (targets.Length != NodeCount)
{
throw new NerotiqException($"target value array length ({targets.Length}) does not match layer size ({NodeCount})");
}
if (_layerTargets == null)
{
try {
_layerTargets = new GpuMatrix((ushort)NodeCount, 1, executionSequence.Context);
} catch (Exception ex)
{
throw new NerotiqException($"Error allocating delta buffer", ex);
}
_layerTargets.SetKernelArg(
_backwardKernel,
11
);
}
_layerTargets.Update(targets, executionSequence);
}
public void SetKernelArgs_Works()
{
var src = @"
__kernel void a(
__global double *data // 0
) {
int global_id = get_global_id(0);
data[global_id] *= data[global_id];
}
";
var srcs = SourceLoader.CreateProgramCollection(src);
var program = Cl.CreateProgramWithSource(
_testScaffold.Context.OpenClContext,
(uint)srcs.Length,
srcs,
null,
out var error
);
error.Should().Be(ErrorCode.Success);
error = Cl.BuildProgram(
program,
1,
new[] { _testScaffold.Context.Device },
string.Empty,
null,
IntPtr.Zero
);
error.Should().Be(ErrorCode.Success);
var kernel = Cl.CreateKernel(program, "a", out error);
error.Should().Be(ErrorCode.Success);
var matrix = new GpuMatrix(3, 3, _testScaffold.Context);
matrix.Update(
new double[] {
1.0, 2.0, 3.0,
4.0, 5.0, 6.0,
7.0, 8.0, 9.0
},
_testScaffold.ExecutionSequence
);
matrix.SetKernelArg(kernel, 0);
_testScaffold.ExecutionSequence.EnqueueNDRangeKernel(
kernel,
1,
null,
new IntPtr [] { new IntPtr(9) },
null
);
using (matrix.Read(_testScaffold.ExecutionSequence))
{
matrix.GetValue(0, 0).Should().Be(1.0);
matrix.GetValue(0, 1).Should().Be(4.0);
matrix.GetValue(0, 2).Should().Be(9.0);
matrix.GetValue(1, 0).Should().Be(16.0);
matrix.GetValue(1, 1).Should().Be(25.0);
matrix.GetValue(1, 2).Should().Be(36.0);
matrix.GetValue(2, 0).Should().Be(49.0);
matrix.GetValue(2, 1).Should().Be(64.0);
matrix.GetValue(2, 2).Should().Be(81.0);
}
}